First Encounter

Before we look at how any of this works, you should experience it once yourself. Ideally right now.

Open an AI assistant, for example ChatGPT at chatgpt.com, Claude at claude.ai, or Gemini at gemini.google.com. The free tier is enough for trying things out anywhere. In front of you sits an empty input field, and what you type into it is called a prompt. That is the technical term for your input, whether it consists of a question, an instruction, or a whole document.

We'll walk through three examples that look identical in the chat window and are still fundamentally different. That is exactly what this chapter is about. I made up the names and numbers in them.

A plan for your money

Start with something that actually concerns you. Type something like this:

“I'm 35, I earn 3,200 euros net, and I have 20,000 euros in my account. I want to build a financial cushion over the next ten years. Make me a concrete plan with a savings rate, an allocation, and the risks. And tell me honestly which parts are unrealistic.”

What comes back is an ordered plan. A possible savings rate, an allocation, a few sentences on the risks, an assessment. Within seconds, based on three numbers.

That doesn't make anything finished. Only you know whether the savings rate fits your life. But you're no longer sitting in front of a blank page, and that is the real difference. Go ahead and ask follow-ups. Ask for a more cautious version. Have it explain what you don't understand. Good use almost always consists of several small steps, not one perfect input.

The same works for a difficult letter, long meeting minutes, a translation, or the outline for a proposal. You describe the task and keep working with the draft.

A sentence that sounds good and is wrong

For the second example, imagine you work in the purchasing department. A supplier has sent a new quote, forty pages of prices, delivery terms, and fine print. By tomorrow you're supposed to say whether you can sign. So you upload the document and have it summarized.

On page 7, two sentences sit close together:

“Invoices are payable within 30 days. Price changes must be announced three months before they take effect.”

In the summary, this becomes:

“The supplier can change prices with 30 days' notice.”

The sentence sounds impeccable. It matches the tone of the document and even names a specific deadline. It is still wrong. The 30 days belong to the payment terms, and price changes take three months.

And now the crucial point: most of the time, this goes fine. As a rule, a model summarizes passages like this correctly, and models get better at it every year. That is exactly where the difficulty lies. Nobody keeps using a tool that is always wrong. With one that is almost always right, at some point you stop checking. And then the one wrong sentence sits in your brief for the management board.

Because you cannot see the difference by looking at the answer. A cautiously worded sentence can be right, a confident one can be off. The form tells you nothing about the content.

Thirty days and three months sit close together in language. In a price negotiation, they are two different worlds. So how carefully you need to check depends not on the task but on the consequences. In a draft for an internal invitation, a mistake may slip through. In a notice period, a loan approval, or a medical recommendation, it may not.

This applies especially to health questions, because an assistant knows neither your examination nor your history. It can help you collect the right questions for your doctor's visit. It is not a diagnosis.

A question that has no right answer

Third example. Type in something that is genuinely on your mind right now:

“Should I take the new job? Twenty percent more pay, but an hour longer commute, and I get along really well with my current boss.”

If you make decisions in a company, the same kind of question looks like this: two quotes for the same machine, one for 84,000 euros with fourteen weeks of delivery time, one for 92,000 euros with five weeks and a supplier you have worked with for years.

What comes back sounds balanced. It sorts the points, names pros and cons, and in the end recommends one of the options. That can be reasonable, and the opposite can be just as reasonable.

Because what's missing here is not knowledge, it's a sense of what matters to you. How much is an hour of your day worth to you? How much does a boss you enjoy working with weigh? How much risk can you carry without it stressing you too much? Nobody can take these questions off your hands, not even a machine with the best arguments. The answer can weight your priorities wrongly or rest on an assumption nobody ever spelled out.

It is useful anyway. Take it as an ordered second opinion, not as an outsourced decision. Because if you follow the recommendation: who actually decided, you or the model?

Suddenly within reach

Three answers, one window, the same font. A usable draft, a convincing false statement, and a recommendation that cannot really be one. That is exactly why the first impression is not enough.

Still, what is happening here is a big deal. The world's knowledge has been sitting on the internet for years, just in a form that rarely helped you. Specialist articles written for specialists. Legal texts written for lawyers. Forums full of answers to questions you never had, or misinformation. Now you ask in your own words and get an explanation the way you understand it, and if you don't understand it, you simply ask again.

For me, these programs have become indispensable in everyday life, when traveling, when planning, for almost every question in between. You can practice a language, have a doctor's letter explained in plain words, learn to program, or work through a quote. All things you used to need an expert or the right specialist knowledge for.

One thing belongs to this: asking follow-up questions. A model doesn't always have all the information your question actually needs, and it can be wrong just the same. You can't see either in the answer. So dig in. Where does this come from? What speaks against it? What if the opposite is true? Asking costs you nothing, an assistant never gets impatient, and you can grill it as long as you want. In a language exercise, a wrong answer doesn't matter. With a contract, a diagnosis, or your savings, you follow up before you act. Exactly this asking is what makes AI so useful as a tool.

What else has a say besides the model

There is an answer in the window, but behind it sits more than the language model alone. The model produces the text. Whether that text is any good depends on four more parts.

1. Context: everything available for this one request. Your prompt, the conversation so far, uploaded files, the application's internal instructions. What isn't in there, and didn't appear in the language model's training, cannot be taken into account. Your history, your priorities, the uneasy feeling about one of the options: all of that stays in your head as long as you don't write it down. And much of it you may not even be able to put fully into words.

2. Tools: other programs the assistant can call. A search, a calculator, access to your documents. In the contract example, a tool could have pulled up the exact passage.

3. Controls: what the system is allowed to do at all. Which data it sees, which action it can trigger, what someone has to approve first. Certain content can be blocked by the operator of the language model, for example.

4. You: You set the task, check the result, and decide whether anything follows from it. A model can write a sentence. It cannot take responsibility for a contract.

Together, these parts form an AI system, and the chat window only shows you its surface.

That sounds like hair-splitting, but it changes how you read a bad answer. Instead of “the AI got it wrong” you can ask which part it was. In the savings plan, the context was missing everything you didn't write down, your rent, say, or a planned house purchase. In the contract, the right passage was in the document, but a tool could have pulled it up directly instead of it being skimmed past in the running text. And in the job decision, nothing in the system was missing, only your sense of what matters. Three very different causes that look the same in the chat window.

We'll look at the complete map with all the parts in Chapter 4. For now, this distinction is enough.

Three questions for the next answer

The three examples turn into three questions. Take them with you the next time an answer sits in front of you:

1. What kind of answer is this? With a draft, you keep working. A claim, you verify. With a judgment, deciding what matters stays with you.
2. What did the system know, and what couldn't it know? What wasn't in the context couldn't flow in, even if it was the decisive part of your question.
3. What happens if the answer is wrong? The bigger the consequence, the more closely you check, and the clearer it has to be who takes responsibility.

The question that stays open

You now have an idea of what these programs can do. A few lines turn into a detailed, plausible text within seconds. But what actually happens when you press send? How does a machine turn three sentences into a plan in fluent language? Is that understanding, or something else entirely?

Before we look at how it works, we take a step back. The story begins long before the first chat window, in the 1950s. It also helps put today's excitement about the topic into perspective.